SVQP 2

Leon Bottou — Sat, 31 Jan 2009 14:22:04 -0000

SVQP1 and SVQP2 are QP solvers for training SVM.

SVQP1 is a solver for convex programming problems with box constraints and at most one equality constraint. It uses the modified gradient projection method (see http://leon.bottou.org/papers/bottou-lin-2006) and can be used as the internal solver in a decomposition loop. This has mostly an historical interest.
SVQP2 is a state-of-the-art SMO solver for quadratic programming problems with box constraints and at most one equality constraint. It performs roughly like LIBSVM or SVMLight, sometimes a bit faster, sometimes a bit slower. It comes as a compact library that is easy to reuse for other algorithms. For instance, the UniverSVM code uses SVQP2 inside.

Both SVQP1 and SVQP2 can be used as standalone c++ libraries or can be exploited using lush bindings. These are not standalone programs.

mloss.org SVQP